A landslide boundary automatic detection method combining SBAS-InSAR and global-local short-term dense connection network
By combining SBAS-InSAR and a global-local short-term dense connection network, the problems of low automation and time-consuming manual drawing in landslide monitoring are solved, and efficient and accurate automatic detection of landslide boundaries and deformation monitoring are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- YUNNAN ELECTRIC POWER TESTING & RES INST (GRP) CO LTD
- Filing Date
- 2025-08-04
- Publication Date
- 2026-07-07
AI Technical Summary
Existing technologies for landslide monitoring suffer from problems such as high labor intensity, low automation, high cost, and low spatial resolution. Furthermore, there is a lack of research on landslide identification that combines deep learning with InSAR, making manual drawing of landslide boundaries time-consuming and labor-intensive.
By employing a combined SBAS-InSAR and global-local short-term dense connection network approach, a global-local short-term dense connection network is established through multi-scene SAR image data preprocessing, geocoding, and data augmentation. This network automatically detects landslide boundaries and integrates time-series deformation, cumulative deformation, and annual average deformation rate.
It achieves efficient and accurate automatic detection of landslide boundaries, and can simultaneously acquire the location, extent, time-series deformation, and deformation rate of landslides, thus improving monitoring efficiency and accuracy.
Smart Images
Figure CN120912904B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of landslide boundary detection technology, and in particular to an automatic landslide boundary detection method that combines SBAS-InSAR and a global-local short-term dense connection network. Background Technology
[0002] Landslides refer to the geological phenomenon of soil and rock masses sliding along a continuous shear failure surface on a slope, caused by human activities and environmental conditions. The Ministry of Natural Resources' "2024 China Natural Resources Bulletin" shows that 5,719 geological disasters occurred nationwide in 2024, of which 3,316 were landslides. Landslide hazard identification and monitoring are crucial for disaster prevention and the protection of people's lives and property.
[0003] When monitoring landslide deformation in high-altitude and mountainous areas, conventional ground monitoring methods such as levels, theodolites, and total stations are often hampered by topography and the natural environment. These methods are characterized by high labor intensity, low efficiency, low automation, and high cost. Furthermore, they can only monitor deformation at certain points, resulting in low spatial resolution and failing to reflect the overall deformation of the landslide. InSAR, a new space-based geodetic observation technology developed in recent decades, later evolved into D-InSAR technology, which has been successfully applied to monitor surface deformation with theoretical accuracy ranging from centimeters to millimeters. Addressing the limitations of D-InSAR technology due to spatiotemporal decoherence and atmospheric phase delay, a series of time-series InSAR technologies, such as SBAS-InSAR, PS-InSAR, and coherent target method, can detect slow surface deformation at the centimeter or even millimeter level over long periods. These technologies offer advantages such as high accuracy, wide monitoring range, low cost, and all-weather operation, and are now widely used in landslide and urban geological hazard monitoring.
[0004] Time-series InSAR technology can detect slow landslide deformations at the centimeter or even millimeter level over long periods, but it requires manually drawing landslide boundaries, which is time-consuming and labor-intensive. The continuous emergence of deep learning methods, such as CNNs and Transformers, has provided a rich technical foundation for intelligent landslide identification. However, compared with the field of computer vision, intelligent processing of InSAR data is still in its early stages, facing the challenge of scarce samples. This is mainly due to the relative difficulty in acquiring InSAR data and the lack of high-quality labeled datasets. Furthermore, there is currently a lack of research on landslide identification that combines deep learning with InSAR. Summary of the Invention
[0005] Therefore, the purpose of this invention is to provide an automatic landslide boundary detection method that combines SBAS-InSAR and global-local short-term dense connection networks, so as to at least solve the above problems.
[0006] The technical solution adopted in this invention is as follows:
[0007] An automatic landslide boundary detection method combining SBAS-InSAR and global-local short-term dense connection networks includes the following steps:
[0008] Step 1: Acquire multi-scene SAR image data, precise orbit data, and DEM data; preprocess the data; and obtain time-series deformation, cumulative deformation, and annual average deformation rate.
[0009] Step 2: Geocode the preprocessed data, create a SAR image landslide dataset, perform data augmentation, and divide it into training, validation, and test sets;
[0010] Step 3: Establish a global-local short-term dense connection network;
[0011] Step 4: Input the SAR image landslide dataset into a global-local short-term dense connection network for training to detect landslide boundaries in SAR images;
[0012] Step 5: Integrate the time-series deformation, cumulative deformation, and annual average deformation rate with the landslide boundary at the detection point to obtain an accurate landslide boundary with time-series deformation, cumulative deformation, and annual average deformation rate.
[0013] Furthermore, in step 1, the data is preprocessed to obtain the time series deformation, cumulative deformation, and annual average deformation rate, specifically as follows:
[0014] Using atmospheric correction methods based on SAR data itself and atmospheric correction methods based on GACOS, the SBAS-InSAR deformation monitoring method is used to perform time-series InSAR processing on multiple SAR images to generate unwrapped D-InSAR phase maps and SAR intensity maps, and to obtain time-series deformation, cumulative deformation and annual average deformation rate.
[0015] Furthermore, step 2 specifically includes:
[0016] Step 21: Use DEM to geocode the unwrapped D-InSAR phase map and SAR intensity map, and fuse them into a three-channel image;
[0017] Step 22: Using ArcGIS to visually interpret the unwrapped D-InSAR phase map, delineate the landslide boundary vector and convert it into raster ground truth labels;
[0018] Step 23: Use Python to crop and normalize the three-channel image and raster ground truth labels, and use methods such as rotation, flipping, and contrast transformation to perform data augmentation;
[0019] Step 24: Divide the enhanced data into training set, validation set and test set to generate SAR image landslide dataset.
[0020] Furthermore, the global-local short-term dense connection network in step 3 includes a short-term dense connection network, a global-local attention module, and a multilayer perceptron.
[0021] Furthermore, the short-term dense connection network consists of 5 stages. Stage 1 and Stage 2 each use a convolutional block, which includes a convolutional layer, a batch normalization layer, and a ReLU activation layer. Stages 3, 4, and 5 each use a short-term dense connection module with a stride of 2 and a short-term dense connection module with a stride of 1.
[0022] Furthermore, the global-local attention module includes a global branch and a local branch, which are used to extract global and local context information, aggregate the extracted global and local context information, and output global-local context information through depthwise convolution, batch normalization and standard convolution; the global branch adopts a window-based multi-head attention mechanism to obtain global context information; the local branch uses two parallel convolutional layers to extract local context information.
[0023] Furthermore, step 4 specifically involves:
[0024] Step 41: Input the SAR image landslide dataset and its raster ground truth labels into a global-local short-term dense connection network. The short-term dense connection network extracts four feature maps. Figure 1-4 The number of output feature channels are 64, 256, 512, and 1024, respectively. Figure 1-4 The sizes are 1 / 4, 1 / 8, 1 / 16 and 1 / 32 of the input image size, respectively;
[0025] Step 42: Input the fourth feature into the convolutional layer and the batch normalization layer, and then input the output feature into the global-local attention module and the multilayer perceptron output feature in sequence. The number of channels of the output feature is 64, and the image size of the output feature is 1 / 32 of the input image.
[0026] Step 43: Perform a weighted summation operation on the output features from Step 42 and the third feature to aggregate the features. Then, input the aggregated features into the global-local attention module and the multilayer perceptron output features in sequence. The number of channels in the output features is 64, and the image size of the output features is 1 / 16 of the input image.
[0027] Step 44: Perform a weighted summation operation on the output features from Step 43 and the second feature to aggregate the features. Input the aggregated features into the global-local attention module and the multilayer perceptron in sequence to output the final features. The number of channels of the final output features is 64, and the image size of the final output features is 1 / 8 of the input image.
[0028] Step 45: Sample and decode the final features to generate the detection results of the landslide boundary;
[0029] Step 46: Based on the final features and ground truth labels, optimize the global-local short-term dense connection network using the Dice loss function and the cross-entropy loss function.
[0030] Compared with the prior art, the beneficial effects of the present invention are:
[0031] This invention provides an automatic landslide boundary detection method that combines SBAS-InSAR and a global-local short-term dense connection network. By establishing a global-local short-term dense connection network, global and local context can be captured at multiple scales, enabling efficient and high-precision automatic extraction of the location and extent of landslides. It can simultaneously acquire landslide boundaries, time-series deformation, cumulative deformation, and annual average deformation rate. Attached Figure Description
[0032] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only preferred embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0033] Figure 1 This is a flowchart of an automatic landslide boundary detection method based on a combination of SBAS-InSAR and a global-local short-term dense connection network, provided by an embodiment of the present invention.
[0034] Figure 2 This is a schematic diagram of a global-local short-term dense connection network provided in an embodiment of the present invention;
[0035] Figure 3 This is a schematic diagram of the global-local attention module provided in an embodiment of the present invention;
[0036] Figure 4 This is a schematic diagram of the Pathfinder-1 strip mode 1 SAR image provided in an embodiment of the present invention;
[0037] Figure 5 This is a schematic diagram of the landslide boundary identification effect of the Road Exploration-1 SAR image provided in an embodiment of the present invention. Detailed Implementation
[0038] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. In the following description, the expression "some embodiments" refers to a subset of all possible embodiments; however, it should be understood that "some embodiments" can be the same subset or different subsets of all possible embodiments and can be combined with each other without conflict.
[0039] In the following description, numerous specific details are set forth in order to provide a more thorough understanding of the invention. However, it will be apparent to those skilled in the art that the invention can be practiced without one or more of these details. In other instances, certain technical features well-known in the art have not been described in order to avoid obscuring the invention.
[0040] It should be understood that the present invention can be embodied in various forms and should not be construed as being limited to the embodiments set forth herein. Rather, providing these embodiments will make the disclosure thorough and complete, and will fully convey the scope of the invention to those skilled in the art. Furthermore, the terminology used herein is intended only to describe particular embodiments and is not intended to limit the invention. When used herein, the singular forms “a,” “an,” and “the” are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the terms “compose” and / or “comprising,” when used in this specification, identify the presence of the stated features, integers, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups. When used herein, the term “and / or” includes any and all combinations of the associated listed items.
[0041] It should also be noted that when an element is referred to as being "fixed to" another element, it can be directly attached to the other element or there may be an intervening element. When an element is referred to as being "connected to" another element, it can be directly connected to the other element or there may be an intervening element. The terms "vertical," "horizontal," "inner," "outer," "left," "right," and similar expressions used herein are for illustrative purposes only and do not represent the only possible implementation.
[0042] To fully understand this invention, a detailed structure will be presented in the following description to illustrate the technical solution proposed by this invention. Optional embodiments of the invention are described in detail below; however, in addition to these detailed descriptions, the invention may have other embodiments.
[0043] Reference Figure 1 , Figure 2 , Figure 3 This invention provides an automatic landslide boundary detection method combining SBAS-InSAR and a global-local short-term dense connection network, comprising the following steps:
[0044] Step 1: Acquire multi-scene SAR image data, precise orbit data, and DEM data; preprocess the data; and obtain time-series deformation, cumulative deformation, and annual average deformation rate.
[0045] The data is preprocessed to obtain the time series deformation, cumulative deformation, and annual average deformation rate, specifically as follows:
[0046] Using atmospheric correction methods based on SAR data itself and atmospheric correction methods based on GACOS, the SBAS-InSAR deformation monitoring method is used to perform time-series InSAR processing on multiple SAR images to generate unwrapped D-InSAR phase maps and SAR intensity maps, and to obtain time-series deformation, cumulative deformation and annual average deformation rate.
[0047] Step 2: Geocode the preprocessed data to create a SAR image landslide dataset, perform data augmentation, and divide it into training, validation, and test sets, specifically including:
[0048] Step 21: Use DEM to geocode the unwrapped D-InSAR phase map and SAR intensity map, and fuse them into a three-channel image;
[0049] Step 22: Using ArcGIS to visually interpret the unwrapped D-InSAR phase map, delineate the landslide boundary vector and convert it into raster ground truth labels;
[0050] Step 23: Use Python to crop the three-channel image and raster ground truth labels to an appropriate size and normalize them, and use methods such as rotation, flipping, and contrast transformation to perform data augmentation;
[0051] Step 24: Divide the enhanced data into training, validation and test sets in a 3:1:1 ratio to generate a SAR image landslide dataset.
[0052] Step 3: Establish a global-local short-term dense connection network, including a short-term dense connection network, a global-local attention module, and a multilayer perceptron;
[0053] The short-term dense connection network consists of 5 stages. Stage 1 and Stage 2 each use a convolutional block, which includes a convolutional layer with a stride of 2, a batch normalization layer, and a ReLU activation layer. Stage 3, Stage 4, and Stage 5 each use a short-term dense connection module with a stride of 2 and a short-term dense connection module with a stride of 1.
[0054] The global-local attention module includes a global branch and a local branch, which are used to extract global and local context information. The extracted global and local context information is aggregated and output as global-local context information after depthwise convolution, batch normalization and standard 1*1 convolution. The global branch uses a window-based multi-head attention mechanism to obtain global context information. The local branch uses two parallel convolutional layers to extract local context information, and adds two batch normalization operations before the summation operation.
[0055] Step 4: Input the SAR image landslide dataset into a global-local short-term dense connection network for training to detect landslide boundaries in SAR images, specifically:
[0056] Step 41: Input the SAR image landslide dataset and its raster ground truth labels into a global-local short-term dense connection network. The short-term dense connection network extracts four feature maps. Figure 1-4 The number of output feature channels are 64, 256, 512, and 1024, respectively. Figure 1-4 The sizes are 1 / 4, 1 / 8, 1 / 16 and 1 / 32 of the input image size, respectively;
[0057] Step 42: Input the fourth feature into the convolutional layer and the batch normalization layer, and then input the output feature into the global-local attention module and the multilayer perceptron output feature in sequence. The number of channels of the output feature is 64, and the image size of the output feature is 1 / 32 of the input image.
[0058] For example, step 42 specifically involves: inputting the fourth feature into a convolutional layer and a batch normalization layer to output the first feature, the number of channels of the output first feature is 64, and the image size of the output first feature is 1 / 32 of the input image;
[0059] The first feature is input into the batch normalization layer, and then input into the global-local attention module to output the second feature. The number of channels of the output second feature is 64, and the image size of the output second feature is 1 / 32 of the input image.
[0060] The second feature is input into the batch normalization layer, and then into the multilayer perceptron to output the third feature. The number of channels of the output third feature is 64, and the image size of the output third feature is 1 / 32 of the input image.
[0061] Step 43: Perform a weighted summation operation on the output features from Step 42 and the third feature to aggregate the features. Then, input the aggregated features into the global-local attention module and the multilayer perceptron output features in sequence. The number of channels in the output features is 64, and the image size of the output features is 1 / 16 of the input image.
[0062] For example, step 43 specifically involves: performing a weighted summation operation on the third feature and the third feature to output a fourth feature, with 64 channels for the output fourth feature and the image size of the output fourth feature being 1 / 16 of the input image;
[0063] The fourth feature is input into the batch normalization layer, and then into the global-local attention module to output the fifth feature. The number of channels of the output fifth feature is 64, and the image size of the output fifth feature is 1 / 16 of the input image.
[0064] The fifth feature is then input into the batch normalization layer again, and then into the multilayer perceptron to output the sixth feature. The number of channels for the output sixth feature is 64, and the image size of the output sixth feature is 1 / 16 of the input image.
[0065] The output features from step 43 are aggregated with the second feature using a weighted summation operation. The aggregated features are then sequentially input into the global-local attention module and the multilayer perceptron to output the final features. The final output features have 64 channels and the image size of the final output features is 1 / 8 of the input image.
[0066] For example, step 44 specifically involves: performing a weighted summation operation on the sixth feature and the second feature to output the seventh feature, which has 64 channels and an image size of 1 / 8 of the input image.
[0067] The seventh feature is input into the batch normalization layer, and then into the global-local attention module to output the eighth feature. The number of channels of the output eighth feature is 64, and the image size of the output eighth feature is 1 / 8 of the input image.
[0068] The eighth feature is input into the batch normalization layer, and then into the multilayer perceptron to output the final feature. The number of channels in the final output feature is 64, and the size of the final output feature image is 1 / 8 of the input image.
[0069] Step 45: Sample and decode the final features to generate the detection results of the landslide boundary.
[0070] Step 46: Based on the final features and ground truth labels, optimize the global-local short-term dense connection network using the Dice loss function and the cross-entropy loss function.
[0071] For example, the DICE loss function is:
[0072]
[0073] in, This represents the number of samples in the batch. The total number of categories, For the sample The true label, For the model to sample The predicted probability output;
[0074] The cross-entropy loss function is:
[0075]
[0076] in, This represents the number of samples in the batch. The total number of categories, For the sample The true label, For the model to sample The predicted probability output represents the sample Category The confidence level.
[0077] Step 5: Integrate the time-series deformation, cumulative deformation, and annual average deformation rate with the landslide boundary at the detection point to obtain an accurate landslide boundary with time-series deformation, cumulative deformation, and annual average deformation rate.
[0078] For example, refer to Figure 4 In this embodiment, Pathfinder-1 SAR imagery was used to monitor a landslide in a certain area. Twenty scenes of domestically produced L-band Pathfinder-1 strip mode 1HH polarimetric SAR imagery and their precise orbital data were selected. ALOS World 3D 30m external DEM data were downloaded from the website (https: / / www.eorc.jaxa.jp / ALOS / en / dataset / aw3d_e.htm), and high spatial resolution zenith pair flow delay data provided by the GACOS system was downloaded from the website (http: / / www.gacos.net). The landslide detection was performed using the method of this invention, and the resulting identification effect image is shown below. Figure 5 As shown, Figure 5 The first column represents the truth label, the second column represents the schematic diagram of the landslide boundary detected by the method of the present invention, and the third column represents the schematic diagram of the landslide boundary detected by the traditional method.
[0079] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for automatic landslide boundary detection combining SBAS-InSAR and global-local short-term dense connection networks, characterized in that, Includes the following steps: Step 1: Acquire multi-scene SAR image data, precise orbit data, and DEM data; preprocess the data; and obtain time-series deformation, cumulative deformation, and annual average deformation rate. Step 2: Geocode the preprocessed data, create a SAR image landslide dataset, perform data augmentation, and divide it into training, validation, and test sets; Step 3: Establish a global-local short-term dense connection network; Step 4: Input the SAR image landslide dataset into a global-local short-term dense connection network for training to detect landslide boundaries in SAR images; Step 5: Integrate the time-series deformation, cumulative deformation, and annual average deformation rate with the landslide boundary at the detection point to obtain an accurate landslide boundary with time-series deformation, cumulative deformation, and annual average deformation rate; The global-local short-term dense connection network in step 3 includes a short-term dense connection network, a global-local attention module, and a multilayer perceptron; The short-term dense connection network consists of 5 stages. Stage 1 and Stage 2 each use a convolutional block, which includes a convolutional layer, a batch normalization layer, and a ReLU activation layer. Stage 3, Stage 4, and Stage 5 each use a short-term dense connection module with a stride of 2 and a short-term dense connection module with a stride of 1. The global-local attention module includes a global branch and a local branch, which are used to extract global and local context information. The extracted global and local context information is aggregated and output as global-local context information after depthwise convolution, batch normalization and standard convolution. The global branch uses a window-based multi-head attention mechanism to obtain global context information. The local branch uses two parallel convolutional layers to extract local context information.
2. The automatic landslide boundary detection method based on a combination of SBAS-InSAR and a global-local short-term dense connection network as described in claim 1, characterized in that, Step 1 involves preprocessing the data and obtaining the time series deformation, cumulative deformation, and annual average deformation rate, specifically as follows: Using atmospheric correction methods based on SAR data itself and atmospheric correction methods based on GACOS, the SBAS-InSAR deformation monitoring method is used to perform time-series InSAR processing on multiple SAR images to generate unwrapped D-InSAR phase maps and SAR intensity maps, and to obtain time-series deformation, cumulative deformation and annual average deformation rate.
3. The automatic landslide boundary detection method based on a combination of SBAS-InSAR and a global-local short-term dense connection network as described in claim 2, characterized in that, Step 2 specifically includes: Step 21: Use DEM to geocode the unwrapped D-InSAR phase map and SAR intensity map, and fuse them into a three-channel image; Step 22: Using ArcGIS to visually interpret the unwrapped D-InSAR phase map, delineate the landslide boundary vector and convert it into raster ground truth labels; Step 23: Use Python to crop and normalize the three-channel image and raster ground truth labels, and use methods such as rotation, flipping, and contrast transformation to perform data augmentation; Step 24: Divide the enhanced data into training set, validation set and test set to generate SAR image landslide dataset.
4. The automatic landslide boundary detection method based on a combination of SBAS-InSAR and a global-local short-term dense connection network as described in claim 3, characterized in that, Step 4 is as follows: Step 41: Input the SAR image landslide dataset and its raster ground truth labels into the global-local short-term dense connection network. Extract four feature maps through the short-term dense connection network. The number of output feature channels of feature maps 1-4 are 64, 256, 512 and 1024 respectively. The sizes of feature maps 1-4 are 1 / 4, 1 / 8, 1 / 16 and 1 / 32 of the input image size respectively. Step 42: Input the fourth feature into the convolutional layer and the batch normalization layer, and then input the output feature into the global-local attention module and the multilayer perceptron output feature in sequence. The number of channels of the output feature is 64, and the image size of the output feature is 1 / 32 of the input image. Step 43: Perform a weighted summation operation on the output features from Step 42 and the third feature to aggregate the features. Then, input the aggregated features into the global-local attention module and the multilayer perceptron output features in sequence. The number of channels in the output features is 64, and the image size of the output features is 1 / 16 of the input image. Step 44: Perform a weighted summation operation on the output features from Step 43 and the second feature to aggregate the features. Input the aggregated features into the global-local attention module and the multilayer perceptron in sequence to output the final features. The number of channels of the final output features is 64, and the image size of the final output features is 1 / 8 of the input image. Step 45: Sample and decode the final features to generate the detection results of the landslide boundary; Step 46: Based on the final features and ground truth labels, optimize the global-local short-term dense connection network using the Dice loss function and the cross-entropy loss function.